Vijayanta Gupta

Essay

The second growth curve

. 10 min read

Fifteen years ago, Marc Andreessen wrote that software was eating the world (Andreessen, WSJ, 2011). The claim landed because it was both audacious and structural. Every industry, he argued, was becoming a software industry, and the businesses that did not see it as such would be eaten from underneath. The thesis was vindicated, in dollars, customers, and market capitalisation, across the decade that followed. Gartner now puts global IT spending at USD 6.31 trillion in 2026, with enterprise software at USD 1.44 trillion growing 15.1 per cent (Gartner, April 2026). The world was eaten. Software won.

The relevant question for B2B software CEOs, chairs, and investors in 2026 is no longer whether software is eating the world. It is what comes next inside software itself. The category that did the eating is now being repriced, recapitalised, and quietly redesigned. The first growth curve of modern enterprise software, the one that ran from on-premise to SaaS, is finishing. A second growth curve, from SaaS to AI-native, has started. It is not a feature cycle. It is a business-model reset, and the operating playbook that powered the first curve is the wrong scaffolding for the second.

The first growth curve rewarded one set of behaviours

The first growth curve had a clean shape. Twenty years ago, Salesforce was a small operation that the press at the time called the ant at the picnic, sized against multi-billion-dollar incumbents in Siebel and Oracle (Cledara, 2024). The economics it brought to the table were brutally simple. Replace a USD 5 million on-premise install, eighteen to twenty-four months to stand up, with a multi-tenant service that started in thirty days and billed per seat. Margins on the operator’s side were extraordinary because each marginal user was effectively free to serve. Boards trained their pattern recognition on the shape: net revenue retention above 110, gross margins around 80, payback inside eighteen months, rule-of-40 in the thirties. Services were kept under a quarter of revenue and used to mask product gaps quietly.

That model ran for a decade and a half. By 2021, public SaaS was growing at thirty per cent and the market was paying for it. By the end of 2025, the median public SaaS revenue growth rate had fallen to 12.2 per cent and valuation multiples had collapsed to 6.4 times ARR, a decade-plus low (SaaS Capital, 2026; Public SaaS Multiples, Q1 2026). The standard reading inside SaaS C-suites is that this is cyclical, the result of a higher rate environment and post-pandemic digestion. The standard reading is wrong. The slowdown is structural, and the cause is now visible.

The hidden break is on the cost line, not the demand line

The break is not on the demand line. Demand for software is still growing. Within the USD 1.44 trillion enterprise software pool, AI is the fastest-growing segment, forecast at USD 2.5 trillion across the broader stack and growing 47 per cent in 2026 (Gartner, January 2026). The break is on the cost line that made SaaS leverage work in the first place. Inference is not free. ICONIQ’s January 2026 snapshot put average AI product gross margin at 52 per cent against the SaaS benchmark of 80, with inference alone consuming roughly twenty-three per cent of revenue at scaling-stage AI B2B companies (ICONIQ, 2026, via Monetizely, 2026). Public SaaS companies that have absorbed AI features into core product are now reporting gross margins ten to seventeen points below pre-AI baselines, with inference cost ratios called out separately in MD&A (SaaS Mag, 2026). The new structural floor for AI-native gross margin is 60 to 70, not 80.

That single change rewires the rest of the operating model. A SaaS company at twenty-five per cent growth and eighty per cent gross margin scored 105 on rule-of-40. The same company at twenty-five per cent growth and sixty-seven per cent gross margin scores 92 (Abacum, 2026). Rule-of-40, payback math, services share, expansion economics, and headcount plans were all calibrated to the eighty-point world. None of them have been recut. The market noticed first. Markets always do.

What the market repriced in February 2026

In February 2026, roughly USD 285 billion of B2B software market capitalisation evaporated in forty-eight hours after Anthropic launched Claude Cowork and demonstrated that AI agents could absorb categories of work previously sold as per-seat software (Taskade, 2026; Financial Content, March 2026). By mid-March, an estimated USD 2 trillion in software market cap had been erased. Morgan Stanley wrote that the era of easy growth for SaaS was over and that enterprises were now reallocating capital toward the AI agents that threaten to replace the very software those agents were supposed to complement (Fortune, February 2026). Jason Lemkin put the operating consequence in one sentence. If ten AI agents can do the work of a hundred reps, the buyer needs ten Salesforce seats, not a hundred (Taskade, 2026).

The reaction was not a panic about AI in the abstract. It was a repricing of a specific operating model. Per-seat pricing, multi-tenant leverage, services-masked product, and SaaS-era underwriting all depend on the buyer paying for human-led work. Once the work itself can be done by software, the unit of value moves. Per-seat pricing fell from twenty-one to fifteen per cent of SaaS in the twelve months to early 2026, hybrid models reached forty-one per cent adoption, and eighty-three per cent of AI-native SaaS now bills on usage or outcome rather than seats (Monetizely, 2026; Stormy AI, 2026).

The new entrants are already operating to the new model. Cursor went from zero to USD 2 billion in annual recurring revenue in roughly three years, the fastest zero-to-two-billion arc on record, ahead of Slack, Zoom, and Snowflake (The Next Web, 2026; Tech Insider, 2026). Harvey reached USD 11 billion in valuation and roughly half of the Am Law 100 inside two years (Tech Insider, 2026, via vertical AI tracking). Sierra crossed USD 150 million in ARR seven quarters after launch (Sierra, 2026). Salesforce Agentforce reached USD 550 million in ARR and 330 per cent growth in months (SaaStr, 2026). None of these companies is selling seats. All of them are selling resolved work.

What AI actually changes in the value chain

Three economic variables move at once. Cost-to-serve goes up because inference is variable and meters with usage. Pricing power shifts because buyers stop paying for access and start paying for outcomes. The unit of value moves from the seat to the resolved task. Each of these is a business-model decision, not a feature decision. None of them is reversible by shipping more AI on the existing roadmap.

The roadmap question is where the most expensive misreads now happen. A list of AI features signals motion to boards and markets while leaving pricing, margin, and the location of value capture untouched. A roadmap is not a strategy. Treating an AI feature backlog as an AI strategy is the AI roadmap theatre that the market is now penalising in valuation. The question underneath the roadmap, the only one that matters, is which line of the profit-and-loss the AI move is built to move first.

The optimistic counterargument inside SaaS C-suites is that inference cost will fall fast enough to restore the eighty-point margin world. Inference cost per token has indeed fallen by roughly an order of magnitude in the past three years, from twenty dollars to forty cents per million tokens for GPT-4-class capability (Introl, 2026). But customer expectations of capability have absorbed the saving. Margin pressure in the businesses that actually ship AI is structural, not transitional (SaaS Mag, 2026).

Cheaper inference is not a margin rescue. It is the price of staying in the market.

What the second curve is doing to capital

The capital response is the cleanest signal. Private equity has read the repricing as opportunity, not threat. Thoma Bravo completed four platform SaaS acquisitions in 2025, including a USD 12.3 billion take-private of Dayforce, and Vista and Thoma Bravo together have deployed more than USD 120 billion in software buyouts since 2019 (Thoma Bravo, 2025; PitchGrade, 2026). Take-private deal flow in late 2025 and Q1 2026 has been the highest since 2021. The trade is straightforward. Buy the asset at a depressed multiple, recut it to the new operating model in private, and hold it through the curve. The investors who already underwrite the second curve are not waiting for the first one to recover.

The board agenda for the second growth curve

Five strategic choices are now on the desk. None is a product decision.

  1. Decide where the unit of value sits, and reprice to it. Seat, usage, outcome, or hybrid. Pick deliberately, and underwrite the gross margin implication of the choice before the next renewal cycle, not after.
  2. Recut the operating model to a sixty-to-seventy-point gross margin world. Rule-of-40 targets, payback assumptions, services share, and headcount plans were calibrated to a regime that no longer exists.
  3. Sequence the reset in the right order. Pricing first, then go-to-market, then product boundaries. Most companies do this in reverse and discover that pricing was the bottleneck after the product change has already shipped.
  4. Replace the AI roadmap review with an AI economics review. The standing board item should track AI-adjusted gross margin, inference cost as a per cent of AI revenue, pricing exposure at next renewal, and the outcome-priced share of new ARR.
  5. Identify the single AI bet that resets the curve. One, not five. The companies winning the second curve are concentrating capital behind one model-level repricing or workflow-collapse move, not spreading it across a feature backlog.

The question that should replace what is our AI roadmap is simpler and harder.

What is our AI-native business model, and which line of the profit-and-loss does it move first.

A CEO who can answer that on one page can run the second growth curve. A CEO who cannot is still running the first.

Andreessen ended his 2011 essay with a claim about investment. The current claim is about operating design. The first curve rewarded leverage in a high-margin distribution model. The second curve rewards economic quality in a high-variable-cost delivery model. Different shape, different scoreboard, different operating playbook. The companies that win the next decade in software will not look like the companies that won the last one. The boards that govern them will not be looking at the same dashboard either.

Sources

AI-native, SaaS, growth, pricing, boards

Back to writing